We consider the problem of classifying a time series as soon as possible, or equivalently classifying a sample with as little computation of features as possible. We formulate the objective of providing a class label as early as possible constrained to guaranteeing with high probability that the early class matches the class that would be assigned to a longer time series (or to a more complete feature set). Results applying the strategy to local quadratic discriminant analysis and to support vector machines on the UCI Time Series datasets show that it can reap substantial time/computation savings in practice. Further computationally efficiency is gained by combining with supervised dimensionality reduction.